Overview
- Pew Research Center’s Data Labs ran about 490,000 English-language pages through the Open Pangram detector and found that a July 2026 random sample showed AI signals on about 10% of pages and about 35% of pages published after ChatGPT’s November 2022 release.
- The study used Common Crawl data and a probabilistic detector that flags both fully AI-generated text and human writing substantially edited with AI, so flagged pages span a spectrum from light edits to full generation.
- AI signals are concentrated on commercial sites, with six-month averages of roughly 9.35% on .com pages, 4.59% on .org, 1.03% on .edu and 0.76% on .gov pages.
- Pew identified rising linguistic fingerprints tied to AI style, including a near doubling of em dash use, a 63% jump in Oxford commas, and higher rates of specific words and negative-parallel phrasing.
- Other teams report different magnitudes — some higher for newly published articles — and experts say measurement limits and growing AI editing features in tools make provenance, disclosure, search ranking, and fact checking the next policy and product priorities.